Speech Recognition Data Cleanup via Homophone Substitution

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Statistical language models in speech recognition require large volumes of training data, which often contain inaccuracies and inconsistencies, affecting performance.

Innovation Solution

An automated data cleanup system using a language model generated from seed text and a dictionary to cleanse training data by applying inverse text normalization and transduction, identifying probabilities of potential substitutions to produce consistent data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If large volumes of training data are used in speech recognition, then the speech recognition performance is improved, but the inaccuracies and inconsistencies in the training data increase, affecting performance

Engineering Contradiction:

Improvevolume of training dataVSAvoidconsistency of training data

Core Design Contradiction:

Quantity of substanceVSReliability

Solution Approach 1:

The system performs preliminary data cleanup and normalization before the training process. By pre-processing the training data to correct inaccuracies and inconsistencies, the system prepares clean data in advance, allowing large volumes of data to be used without compromising quality. This resolves the contradiction by addressing data quality issues before they can affect training performance.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system employs feedback mechanisms where the cleaned and normalized data is used to train language models, which then provide feedback on data quality and consistency. This iterative process allows continuous improvement of data quality while maintaining large data volumes, resolving the contradiction between quantity and reliability.

Inventive Principle:

Principle #23Feedback

2Quantity of substance

If training data contains inaccuracies and inconsistencies, then the data volume can be large, but the speech recognition performance deteriorates

Engineering Contradiction:

Improvevolume of training dataVSAvoidspeech recognition performance

Core Design Contradiction:

Quantity of substanceVSMeasurement precision

Solution Approach 1:

The system performs preliminary cleaning and normalization operations on training data before it is used for training speech recognition models. This pre-processing step removes inaccuracies and inconsistencies, ensuring that only high-quality data is used for training, thereby maintaining both large data volume and high performance.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system introduces intermediate processing steps including data cleaning modules, normalization layers, and quality filtering mechanisms that act as intermediaries between raw data and training models. These intermediaries filter out poor quality data while preserving valuable data, enabling large volumes of data to be processed without sacrificing performance.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Reliability

If automated data cleanup is applied to increase consistency, then the quality of training data is improved, but the processing time and complexity increase

Engineering Contradiction:

Improveconsistency of training dataVSAvoiddata processing time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system performs data cleaning and normalization as preliminary actions during the data collection phase, rather than as post-processing steps. By addressing data quality issues early in the pipeline, the system reduces the need for time-consuming cleanup operations later, thereby improving consistency without significantly increasing overall processing time.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system implements continuous data cleaning and normalization processes that operate parallel to data collection and model training. This continuous approach ensures data quality is maintained throughout the entire pipeline, reducing the need for separate batch processing steps and minimizing time loss.

Inventive Principle:

Principle #20Continuity of useful action

Data Source

PatentUS9460708B2Automated data cleanup by substitution of words of the same pronunciation and different spelling in speech recognition

Publication Date: 2016.10.04 MICROSOFT TECHNOLOGY LICENSING LLC

AI summary

The described implementations relate to automated data cleanup. One system includes a language model generated from language model seed text and a dictionary of possible data substitutions. This system also includes a transducer configured to cleanse a corpus utilizing the language model and the dictionary. The transducer can process speech recognition data in some cases by substituting a second word for a first word which shares pronunciation with the first word but is spelled differently. In some cases, this can be accomplished by establishing corresponding probabilities of the first word and second word based on a third word that appears in sequence with the first word.